On November 20, Anthropic mathematician Levent Alpöge announced on his personal website that he and Claude had found a proof of the Riemann Hypothesis. A few hours later, Ravi Vakil, a Stanford mathematics professor and president of the American Mathematical Society, posted a different draft on his blog. Its author is Alessandro Masullo, and it targets the broader Generalized Riemann Hypothesis. Vakil specifically noted that no AI was used at any stage.
A problem posed in 1859, with a $1 million prize attached, produced two unrelated proofs on the same day. So far, neither has been accepted by the mathematical community.
250,000 pages, released as code only
Alpöge left himself some room in his wording:
I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.
He says the method also covers the Generalized Riemann Hypothesis. According to reports from The Decoder and Scientific American, the proof would exceed 250,000 pages if printed, so he released only the formalized version written in Lean, split into 14 data packages. Lean is a proof assistant language in which a program checks every step of a derivation against the axioms.
The approach builds on a technique the mathematician Stepanov used on a different class of problems: derivatives generally do not work over the integers, so you imitate them with a family of special transformations. As the reports quote Alpöge, the key step came from Claude: for a sequence of length X, the number of chained transformations should be the square root of X. He spent several months setting the direction on simplified toy models first; once the work moved to the real problem, the AI left him "far behind" in the second half.
Alpöge also admits that in some subfields Claude went further than he did, and that he has not yet fully absorbed the surrounding mathematics behind the results he is handing over. He plans to write a human-readable version over the next few months. None of the reports say which Claude version was used or how much compute it consumed.
Who has checked it
The advantage of a formal proof is that the derivation does not rely on human eyes. It leaves one gap, though: if the statement written into the code at the very start does not match the Riemann Hypothesis itself, the program cannot catch that. A person has to check this step, and the reports say the workload is small.
Accounts of the verification progress differ. Nature's report on November 20 said no independent check existed at that point. An updated Scientific American article says MIT number theorist Andrew Sutherland has independently compiled the code and called it "a victory for humanity." James Maynard of the University of Oxford told Nature he expects the proof to be correct, provided it "passes full verification." Andrew Granville of the University of Montreal described Alpöge's explanatory documentation as "a mess." Tristan Buckmaster of New York University said the proof is complicated and will take a long time to digest.
Anthropic has issued no company-level announcement. Nature said the company did not immediately respond to a request for comment. All that can be confirmed is Alpöge's own account: he obtained the proof late on the Sunday night before the announcement, then asked CEO Dario Amodei and got his agreement before going public. The announcement landed the day after Anthropic's public listing; he says the timing was unfortunate and denies it was a PR move.
The other draft was flagged about eight hours later
Masullo's manuscript is titled "The Music of the Primes." Vakil wrote on his blog that Stanford number theorist Brian Conrad had been reviewing it for several months. According to the reports, Masullo had planned to spend another year polishing it at Stanford before going public, and the news from Anthropic's side pushed them to release it early.
About eight hours later, Vakil added an update: several people had reported problems in the manuscript, some of them found with the help of AI models. He wrote that every problem must be resolved to the satisfaction of peers, and that he takes no position on whether the proof strategy is correct. Conrad told Scientific American that the paper is undergoing "major revisions." The blog does not say which problems they are or whether they can be fixed.
From the Jacobian Conjecture to the Riemann Hypothesis
Looking back over the past half year, Alpöge and Claude have appeared on the same thread several times:
- In July, Alpöge published a solution to the Jacobian Conjecture generated by Claude;
- In September, Buckmaster publicly suggested that OpenAI may have drawn on earlier work by him and Alpöge for the Navier–Stokes problem;
- In October, OpenAI released a batch of results on primes, and Alpöge said a key idea of his was inspired by them;
- In November, OpenAI announced a solution to the Navier–Stokes Millennium Problem, and a few days later came this one.
These events share one trait: the result first appears in machine-checkable form, and the human-readable version comes later. A peer-review reform proposal from University of Chicago mathematicians earlier on, and what Terence Tao calls a "proof glut," both address how peer review can keep up in such a situation.
The prize is still a long way from being paid out. The Clay Mathematics Institute's Millennium Prize rules require a proof to be published in a qualifying journal, followed by at least two years of waiting and general acceptance by the mathematical community. How a proof that exists only as Lean code, with no paper yet, would go through that process is unclear, and the Clay Institute has not been seen commenting on it. There is likewise no timetable for when the revisions to Masullo's draft will be finished.
Sources: The Decoder, Scientific American, CocoLoop, Nature News, Ravi Vakil's blog; The Decoder was used to check the number of Lean packages and the page-count basis, and Vakil's blog to check the draft's update and his stated position.